Network Sparsity and Persona Diversity Drive Problem-Solving Capabilities of Agentic AI Systems

This thesis explores the performances and the dynamics that emerge from interactions among complex networks of Large Language Models (LLMs). By generating debates between ensembles of LLM-based agents assigned to specific roles (personas) and allowing them to interact freely within an unfixed topology, we analyze the resulting collective abilities and the patterns of interactions among speaker-receiver pairs, to investigate whether personas can drive LLMs’ conversational preferences when facing specific problems. This novel approach to multi-agent debate is called DynMAD (Dynamic Multi-Agent Debate) and unifies several key strengths of state-of-the-art findings in this field. One of the main objectives is to understand how the formation of complex communication networks in DynMAD can impact the conversational dynamics and the accuracy on a given task, and which metrics are more relevant for characterizing agents’ behavior. Additionally, this thesis investigates whether assigning diverse personas to agents involved in complex problem-solving tasks, without an underlying communication structure, can provide insights into collective intelligence phenomena, such as the "wisdom of crowds," where groups of non-expert problem solvers are able to outperform individual experts. The results may offer new perspectives on LLMs’ collective behavior and their potential impact on complex network-based applications in social sciences and other fields, presenting also a novel approach to important topics such as reducing computational costs and improving content factuality in multi-agent systems.